A tool wear state detection method for a numerical control machine tool

Abstract

The application discloses a tool wear state detection method for a numerical control machine tool, and particularly relates to the technical field of numerical control machine tool processing state monitoring, and comprises the following steps: synchronously sampling a sensor signal on a spindle, extracting a transient response component through time-frequency conversion, and fusing and coding into a multi-source sensor transient characteristic map; calculating a working condition regularization weight factor according to current cutting parameters, executing a gated recurrent filtering on the map to enhance a slow-varying trend, and obtaining a weighted characteristic flow; projecting the weighted characteristic flow to two orthogonal one-dimensional subspaces which are sensitive to wear and sensitive to load in parallel, and outputting a wear trend indicator; the application realizes real-time calculation of a working condition regularization weight and recursive forgetting filtering to inhibit load mutation interference, utilizes wear-load orthogonal projection decomposition to make the wear trend not affected by sudden disturbance, and performs single-step bias correction on a neural network output layer to realize rapid detection after sudden disturbance with minimum calculation cost and low labeled data dependence.

Description

Technical Field

[0001] This invention relates to the field of CNC machine tool machining condition monitoring technology, and more specifically, to a method for detecting tool wear condition in CNC machine tools. Background Technology

[0002] During the cutting process of CNC machine tools, the continuous friction and impact between the cutting tool and the workpiece inevitably lead to tool wear. When the tool wear accumulates to a certain extent, it will cause problems such as increased cutting force, decreased machining accuracy and deterioration of workpiece surface quality. In severe cases, it may even cause tool breakage and damage to the machine tool spindle. Therefore, real-time detection of tool wear during the machining process has important engineering application value.

[0003] In recent years, considering the frequent changes in process parameters such as spindle speed, feed rate, and depth of cut during actual machining, various dynamic modeling schemes have been developed to address the interference of changing working conditions on wear detection. For example, multi-task learning models can be trained by collecting sample data under multiple working conditions, enabling the model to adapt to various conditions simultaneously; or transfer learning can be used to adjust the parameters of a model trained under one working condition to another; or domain adaptation methods can be employed to align the feature distributions under different working conditions to the same common feature space. These methods, within the known range of working conditions or under predictable working condition variations, have achieved a certain degree of adaptive detection, providing effective technical support for tool wear monitoring in dynamic machining scenarios.

[0004] However, in engineering practice, there are still machining scenarios that existing dynamic modeling schemes struggle to effectively handle. When the cutting tool encounters sudden changes in workpiece material hardness, uneven allowance distribution leading to instantaneous heavy-load impacts, or a rapid, nonlinear increase in cutting force due to accelerated tool wear, the aforementioned models based on transfer learning or domain adaptation need to update their parameters online during the detection process to cope with the abrupt signal distribution. The iterative computation time required for online model updates is typically hundreds of milliseconds to several seconds. During this period, the machine tool may have already completed dozens to hundreds of cutting impacts, making it impossible to respond to and discriminate sudden signal jumps in real time, resulting in significant deviations in wear condition assessment during this time period. Furthermore, constructing a priori labeled dataset covering all possible abrupt changes requires substantial experimental and time costs, limiting the feasibility and economic viability of large-scale deployment of the aforementioned dynamic modeling schemes in industrial settings.

[0005] Therefore, how to achieve real-time and accurate detection of tool wear under dynamic changes in cutting parameters and sudden disturbances while maintaining low dependence on labeled data and online computing overhead is a technical problem that needs to be further solved by existing technologies. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting tool wear condition in CNC machine tools includes the following steps: The first step is to synchronously sample the raw sensing signals output by the force sensor, vibration sensor and acoustic emission sensor installed on the spindle of the CNC machine tool, and to perform short-time Fourier transform and wavelet packet decomposition on the signals of each sensor to extract the transient response components of each sensor signal in the time domain, frequency domain and time-frequency domain. All transient response components are fused and encoded into a multi-source sensing transient feature map in the order of channels. The second step is to obtain the cutting parameters of the CNC machine tool currently in operation. These cutting parameters include at least the spindle speed, feed rate, and depth of cut. Based on the cutting parameters, the working condition regularization weight factor is calculated. The working condition regularization weight factor is used to perform a gated cyclic filtering operation on the multi-source sensor transient feature map. This operation selectively enhances the slowly changing trend components in the multi-source sensor transient feature map that are positively correlated with the cumulative tool wear, while suppressing the high-frequency interference components caused by sudden load changes during the cutting process, thereby obtaining the weighted feature flow after working condition normalization. The third step is to map the weighted feature flow into two mutually orthogonal one-dimensional feature subspaces spanned by the first and second projection directions through parallel projection. The first subspace corresponding to the first projection direction is constructed to be sensitive to tool wear, and the second subspace corresponding to the second projection direction is constructed to be sensitive to instantaneous cutting load. The instantaneous cutting load component is decomposed from the weighted feature flow through this parallel projection and then removed, while retaining and outputting a pure wear trend indicator. The fourth step involves inputting the wear trend indicator into the wear state assessment neural network that has been pre-established and stored in the CNC system. The statistical deviation between the wear trend indicator and the sample distribution under each known working condition category used during the training of the neural network is calculated. When the statistical deviation exceeds the pre-set deviation threshold, the current prediction confidence is determined to have decreased, and the neural network is triggered to perform online bias correction on its output value. After correction, the current tool wear state value is obtained, and this value is output as the final detection result.

[0007] In a preferred embodiment, the specific process of synchronous sampling is as follows: The three sensor signals are simultaneously converted from analog to digital at a preset sampling frequency to obtain three digitized raw timing signals. Perform a short-time Fourier transform on each of the original digitized time-series signals to extract the transient amplitude sequence in the time domain and the transient spectral energy distribution sequence in the frequency domain for each signal. Wavelet packet decomposition is performed on each digitized original time-series signal, and the reconstruction coefficients of each frequency band node after decomposition are used as the transient response components of the signal in the time-frequency domain. The time-domain amplitude sequence, frequency-domain energy distribution sequence, and time-frequency domain reconstruction coefficients of the force sensor channel are combined to form the feature vector of the channel. The corresponding three sequences of the vibration sensor channel are combined to form the feature vector of the channel. The corresponding three sequences of the acoustic emission sensor channel are combined to form the feature vector of the channel, resulting in three feature vectors.

[0008] In a preferred embodiment, after obtaining three feature vectors, canonical correlation analysis is performed on the three feature vectors to solve for a set of common projection directions. Each direction in the set of common projection directions corresponds to a correlation coefficient. The correlation coefficient characterizes the overall correlation between the force sensor feature vector and the vibration sensor feature vector, and between the vibration sensor feature vector and the acoustic emission sensor feature vector. When the number of common projection directions in the set is not less than two, the direction corresponding to the one with the largest correlation coefficient is selected as the first common projection direction, and the direction corresponding to the one with the second largest correlation coefficient is selected as the second common projection direction. The three feature vectors are projected onto the two-dimensional correlation subspace spanned by the first common projection direction and the second common projection direction, respectively, to obtain three projection vectors with the same dimension. The columns of the projection vectors are arranged in the order of sampling time. The three projection vectors are arranged longitudinally in the order of the force sensor, vibration sensor, and acoustic emission sensor channels to form a two-dimensional projection matrix with three channel rows and fixed dimension columns. This two-dimensional projection matrix is ​​the multi-source sensing transient feature map.

[0009] In a preferred embodiment, the specific process for calculating the operating condition regularization weight factor is as follows: The spindle speed, feed rate, and depth of cut are read in real time from the shared memory area of ​​the CNC system. At the same time, the rated spindle speed, rated feed rate, and rated depth of cut are read from the preset CNC system. The rated spindle speed, rated feed rate, and rated depth of cut are all positive numbers greater than zero. Divide the spindle speed value by the rated spindle speed value to obtain the speed normalization factor; divide the feed rate value by the rated feed rate value to obtain the feed normalization factor; divide the depth of cut value by the rated depth of cut value to obtain the depth of cut normalization factor; multiply the speed normalization factor, feed normalization factor, and depth of cut normalization factor, and use the result of the multiplication operation as the working condition regularization weight factor.

[0010] In a preferred embodiment, the gated cyclic filtering operation is specifically a weighted summation filtering based on the recursive forgetting coefficient. The specific process of performing the gated cyclic filtering operation on the multi-source sensor transient feature map using the operating condition regularization weight factor is as follows: The current feature vector in the multi-source sensor transient feature map is multiplied element-wise with the operating condition regularization weight factor to obtain the gated current feature vector. The gated current feature vector is then weighted and summed with the filtered historical feature vector from the previous time step according to a preset recursive forgetting coefficient, where the weight of the historical feature vector is greater than the weight of the gated current feature vector. The result of the weighted sum is used as the filtered output vector for the current time step. The filtered output vectors corresponding to all sampling times are arranged vertically in chronological order of acquisition time to form a weighted feature stream that has been normalized by the operating condition.

[0011] In a preferred embodiment, the specific process of mapping the weighted feature stream to two mutually orthogonal one-dimensional feature subspaces spanned by the first and second projection directions through parallel projection is as follows: During the offline calibration phase, historical weighted feature flow samples of the CNC machine tool under various cutting conditions are collected. For each historical sample, the corresponding true value of tool wear and the true value of instantaneous cutting load are simultaneously labeled. All historical samples labeled with the true value of tool wear constitute a first sample set, and the mean vector of this first sample set is calculated. This mean vector is used as the wear-sensitive reference center. All historical samples labeled with the true value of instantaneous cutting load constitute a second sample set, and the mean vector of this second sample set is calculated. This mean vector is used as the load-sensitive reference center. A connection vector is constructed with the wear-sensitive reference center as the starting point and the load-sensitive reference center as the ending point. This connection vector is used as the first projection direction. When the wear-sensitive reference center is not a zero vector and the connection vector is not linearly correlated with the wear-sensitive reference center, the orthogonal complement vector of the connection vector in the linear space where the wear-sensitive reference center is located is calculated. This orthogonal complement vector is used as the second projection direction. The first projection direction and the second projection direction are used as two mutually orthogonal projection bases for parallel projection.

[0012] In a preferred embodiment, the specific process of retaining and outputting the pure wear trend indication is as follows: The weighted feature stream at the current moment is projected onto the first projection direction and the second projection direction respectively to obtain the first projection coefficient along the first projection direction and the second projection coefficient along the second projection direction; the first projection coefficient is multiplied by the first projection direction to reconstruct the wear-sensitive feature component at the current moment; the second projection coefficient is multiplied by the second projection direction to reconstruct the load-sensitive feature component at the current moment. An orthogonal complementary projection operator is constructed based on the first and second projection directions. The orthogonal complementary projection operator is a reprojection operator along the first projection direction in the remaining space after removing the load-sensitive feature components from the weighted feature flow. The wear-related feature vector output after the orthogonal complementary projection operator is the wear trend indicator at the current moment. The wear trend indicators at all sampling moments are arranged in chronological order of acquisition time and then output.

[0013] In a preferred embodiment, the sample distribution is characterized by the sample center vector and sample covariance matrix under each working condition category. The specific process for calculating the statistical deviation between the wear trend indicator and the sample distribution under each known working condition category used during neural network training is as follows: Read the sample center vectors and sample covariance matrices of the wear condition assessment neural network stored in the training phase for each known working condition category from the CNC system. The number of training samples for each known working condition category is greater than the feature vector dimension of that category. The current wear trend indicator is compared with the sample center vector of each known working condition category to obtain the deviation vector of the wear trend indicator relative to each known working condition category. For each known working condition category, the deviation vector corresponding to the category is subjected to a quadratic form operation with the inverse matrix of the sample covariance matrix corresponding to the category to obtain the Mahalanobis distance value of the category. The Mahalanobis distance values ​​corresponding to all known working condition categories are weighted and summed, where the weight of each Mahalanobis distance value is the proportion of the number of training samples in that category to the total number of training samples. The result of the weighted sum is used as the statistical deviation between the wear trend indicator and the sample distribution of each known working condition category.

[0014] In a preferred embodiment, when the statistical deviation exceeds a preset deviation threshold, the neural network is triggered to perform online bias correction on its output value. The specific process for obtaining the current tool wear state value after correction is as follows: The statistical deviation is compared with the deviation threshold pre-stored in the CNC system. When the statistical deviation is less than or equal to the deviation threshold, the output value of the wear state assessment neural network at the current moment is directly used as the current tool wear state value. When the statistical deviation exceeds the deviation threshold, the current bias term of the output layer of the wear state assessment neural network is obtained, along with the weight matrix of the output layer. The current wear trend indicator is used as the input to the neural network, and after transformation by the hidden layer, the input feature vector of the output layer is obtained. This input feature vector is multiplied by the weight matrix and then added to the current bias term to obtain the current predicted wear value. The arithmetic mean of each component of the wear trend indicator is taken as the expected wear value, and the difference between the predicted wear value and the expected wear value is calculated. A step size factor is preset, with a value between 0.01 and 0.1. The difference is multiplied by the step size factor and accumulated into the current bias term to obtain the updated bias term. The updated bias term replaces the current bias term of the output layer of the wear state assessment neural network, and the output value of the output layer is recalculated using the replaced network. This recalculated output value is then output as the current tool wear state value.

[0015] The technical effects and advantages of this invention are as follows: This invention simultaneously extracts the time-domain transient amplitude sequence, frequency-domain spectral energy distribution sequence, and time-frequency domain wavelet packet reconstruction coefficient sequence from the raw signals of force sensors, vibration sensors, and acoustic emission sensors. All transient response components are then fused and encoded into a multi-source sensor transient feature map, providing a multi-dimensional information foundation covering the time, frequency, and time-frequency domains for tool wear condition detection. This feature map preserves the complete correlation structure of the three physical quantity sensor signals in both the time and frequency band dimensions at the same sampling time, effectively avoiding the problem of information loss or insufficient representational ability of single sensor or single-domain features during sudden changes in operating conditions. This lays a reliable data foundation for subsequent adaptive processing of operating conditions and accurate extraction of wear features.

[0016] This invention calculates a working condition regularization weighting factor in real time based on the current spindle speed, feed rate, and depth of cut. This weighting factor is then used to perform gated cyclic filtering on the transient feature maps from multiple sources. This allows for real-time mapping of the feature map amplitude to a unified reference scale without iteratively updating model parameters. Simultaneously, a recursive forgetting mechanism, where historical weights are greater than current weights, suppresses high-frequency interference caused by sudden changes in cutting load. Furthermore, by projecting the filtered weighted feature stream in parallel onto two orthogonal one-dimensional subspaces sensitive to wear and load, the mathematical separation of wear characteristics and instantaneous load disturbances at the signal level is fundamentally achieved. This ensures that even if sudden changes in material hardness or heavy-load impacts occur during machining, the wear trend indicator will not shift due to instantaneous jumps in the load signal. This solves the problem of existing dynamic modeling methods being unable to respond promptly to sudden disturbances due to online update delays.

[0017] This invention processes wear trend indicators by pre-establishing and storing a wear state evaluation neural network in the CNC system, and calculates the statistical deviation between the indicator and the sample distribution of each known working condition category as the confidence criterion. A single-step bias correction of the neural network output layer is triggered only when the statistical deviation exceeds a pre-set threshold, without retraining the entire network or adjusting connection weights. This mechanism enables rapid compensation for output drift under unknown working conditions through the cumulative update of the output layer bias term within a single sampling period. Furthermore, it avoids the enormous cost of collecting labeled datasets across the entire working condition range, thus achieving rapid wear detection after sudden disturbances under limited labeled data conditions. Attached Figure Description

[0018] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of a tool wear condition detection method for CNC machine tools according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1 The following examples were obtained: Example 1: A method for detecting tool wear condition in CNC machine tools, comprising the following steps: The first step is to synchronously sample the raw sensing signals output by the force sensor, vibration sensor and acoustic emission sensor installed on the spindle of the CNC machine tool, and to perform short-time Fourier transform and wavelet packet decomposition on the signals of each sensor to extract the transient response components of each sensor signal in the time domain, frequency domain and time-frequency domain. All transient response components are fused and encoded into a multi-source sensing transient feature map in the order of channels. The second step is to obtain the cutting parameters of the CNC machine tool currently in operation. These cutting parameters include at least the spindle speed, feed rate, and depth of cut. Based on the cutting parameters, the working condition regularization weight factor is calculated. The working condition regularization weight factor is used to perform a gated cyclic filtering operation on the multi-source sensor transient feature map. This operation selectively enhances the slowly changing trend components in the multi-source sensor transient feature map that are positively correlated with the cumulative tool wear, while suppressing the high-frequency interference components caused by sudden load changes during the cutting process, thereby obtaining the weighted feature flow after working condition normalization. The third step is to map the weighted feature flow into two mutually orthogonal one-dimensional feature subspaces spanned by the first and second projection directions through parallel projection. The first subspace corresponding to the first projection direction is constructed to be sensitive to tool wear, and the second subspace corresponding to the second projection direction is constructed to be sensitive to instantaneous cutting load. The instantaneous cutting load component is decomposed from the weighted feature flow through this parallel projection and then removed, while retaining and outputting a pure wear trend indicator. The fourth step involves inputting the wear trend indicator into the wear state assessment neural network that has been pre-established and stored in the CNC system. The statistical deviation between the wear trend indicator and the sample distribution under each known working condition category used during the training of the neural network is calculated. When the statistical deviation exceeds the pre-set deviation threshold, the current prediction confidence is determined to have decreased, and the neural network is triggered to perform online bias correction on its output value. After correction, the current tool wear state value is obtained, and this value is output as the final detection result.

[0021] The specific process for synchronously sampling the raw sensing signals output by the force sensor, vibration sensor, and acoustic emission sensor is as follows: The three sensor signals are simultaneously converted from analog to digital at a preset sampling frequency to obtain three digitized raw timing signals. The sampling frequency is set according to the Nyquist sampling theorem, which states that the sampling frequency should not be less than twice the highest frequency component of the sensor signal, while also taking into account the maximum frequency limit supported by the CNC system's data acquisition card. The sampling frequency of the force sensor and vibration sensor channels can be selected as 2000 Hz, and the sampling frequency of the acoustic emission sensor channel can be selected as 500 kHz. The effective frequency range for the stress sensor and vibration sensor signals is 0 to 1000 Hz, and the effective frequency range for the acoustic emission sensor signals is 0 to 250 kHz. This combination covers the main energy concentration frequency bands of the force signal, vibration signal, and acoustic emission signal during tool cutting.

[0022] For each digitized original time-series signal, a short-time Fourier transform (SFT) is performed to extract the transient amplitude sequence in the time domain and the transient spectral energy distribution sequence in the frequency domain. The SFT uses a Hanning window to segment the original time-series signal, with the window length set according to the signal's stationary interval. The Hanning window length is typically set to 128 sampling points, with a 50% window overlap. Therefore, the transient amplitude sequence in the time domain is the sequence obtained by arranging the maximum signal amplitude within each window in sliding order, and the transient spectral energy distribution sequence in the frequency domain is the sequence obtained by arranging the energy amplitude of each frequency component after the Fourier transform within each window in ascending order of frequency.

[0023] Wavelet packet decomposition is performed on each channel of the original digitized time-series signal. The reconstruction coefficients of each frequency band node after decomposition are used as the transient response components of that signal in the time-frequency domain. With a decomposition level of 3, 8 frequency band nodes are obtained, arranged sequentially from low to high frequency. The reconstruction coefficients of each frequency band node represent a sequence of signal amplitude changes over time within that band, reflecting the time-varying energy distribution of the original signal within the corresponding band. The wavelet basis function used for wavelet packet decomposition is the db4 wavelet from the Daubechies family, which possesses good compact support and regularity, making it suitable for processing non-stationary impact signals generated during tool wear.

[0024] The time-domain amplitude sequence, frequency-domain energy distribution sequence, and time-frequency domain reconstruction coefficients of the force sensor channel are combined to form the feature vector of the force sensor channel. The corresponding three sequences of the vibration sensor channel are combined to form the feature vector of the vibration sensor channel. The corresponding three sequences of the acoustic emission sensor channel are combined to form the feature vector of the acoustic emission sensor channel, resulting in three feature vectors.

[0025] The total dimension of the feature vector for each channel is the sum of the length of the time-domain amplitude sequence, the length of the frequency-domain energy distribution sequence, and the length of the time-frequency domain reconstruction coefficient sequence. For the case where the length of the time-domain amplitude sequence is N1, the length of the frequency-domain energy distribution sequence is N2, and the length of the time-frequency domain reconstruction coefficient sequence is N3, then the dimension of the feature vector for each channel is N1 plus N2 plus N3, and the feature vectors of the three channels have the same dimension.

[0026] After obtaining three feature vectors, canonical correlation analysis is performed on them to solve for a set of common projection directions. Each direction in this set corresponds to a correlation coefficient, which characterizes the overall correlation between the force sensor feature vector and the vibration sensor feature vector, and between the vibration sensor feature vector and the acoustic emission sensor feature vector. When there are at least two such common projection directions, the direction with the largest correlation coefficient is selected as the first common projection direction, and the direction with the second largest correlation coefficient is selected as the second common projection direction. Canonical correlation analysis obtains multiple common projection directions and their corresponding correlation coefficients by solving for two sets of canonical correlation variables between the force sensor feature vector and the vibration sensor feature vector, and between the vibration sensor feature vector and the acoustic emission sensor feature vector.

[0027] The correlation coefficient ranges from 0 to 1. A larger value indicates a stronger linear correlation between the two feature vectors in the projection direction. If the number of common projection directions is M, the correlation coefficients are arranged from largest to smallest as r1 ≥ r2 ≥ ... ≥ rM. The direction corresponding to r1 is selected as the first common projection direction, and the direction corresponding to r2 is selected as the second common projection direction. M is required to be no less than 2.

[0028] If five common projection directions are obtained, with correlation coefficients of 0.92, 0.85, 0.73, 0.61, and 0.48 respectively, then the direction corresponding to the largest correlation coefficient (0.92) is the first common projection direction, and the direction corresponding to the second largest correlation coefficient (0.85) is the second common projection direction. The three feature vectors are projected onto the two-dimensional correlation subspace spanned by the first and second common projection directions, resulting in three projection vectors with the same dimension. The columns of the projection vectors are arranged in chronological order of sampling time.

[0029] The projection operation of a 3D feature vector onto a 2D correlated subspace is as follows: Each channel's feature vector is inner-productted with the first and second common projection directions to obtain two projected coordinate values ​​for that channel in the two-dimensional correlation subspace. These two projected coordinate values ​​from all sampling times are then arranged chronologically to form the channel's projection vector. Each channel's projection vector has the same number of columns, equal to the total number of sampling times, and each column corresponds to the two projected coordinate values ​​for that channel at the same sampling time.

[0030] The three projection vectors are arranged vertically according to the channel order of the force sensor, vibration sensor, and acoustic emission sensor, forming a two-dimensional projection matrix with three channel rows and fixed-dimensional columns. This two-dimensional projection matrix is ​​the transient feature map of the multi-source sensing system. Each row of the two-dimensional projection matrix corresponds to the complete sequence of projected coordinates for one sensor channel, and each column corresponds to the combination of projected coordinates for the three sensor channels at the same sampling time. The matrix has 3 rows and the number of columns is equal to the total number of sampling times. This two-dimensional projection matrix serves as the standardized input basis for subsequent gated cyclic filtering operations. Its row direction represents the sensor channel dimension, and its column direction represents the time evolution dimension.

[0031] The specific process for calculating the operating condition regularization weight factor is as follows: The spindle speed, feed rate, and depth of cut are read in real time from the shared memory area of ​​the CNC system. Simultaneously, the rated spindle speed, rated feed rate, and rated depth of cut are also read from the preset values ​​in the CNC system. All three values ​​are positive numbers greater than zero. The spindle speed is obtained from the spindle encoder signal of the CNC system, the feed rate from the feed servo driver, and the depth of cut from the preset values ​​in the machining program or tool compensation values. All three are the actual working parameters of the machine tool at the current operating moment. The rated spindle speed, rated feed rate, and rated depth of cut are stored in the parameter configuration area of ​​the CNC system, and their values ​​are set according to the equipment manual provided by the machine tool manufacturer. For example, for a machining center with a rated spindle speed of 10,000 rpm, the rated feed rate is typically 0.2 mm / rpm, and the rated depth of cut is typically 1.5 mm. These values ​​represent the standard operating points under the machine tool's design conditions.

[0032] Divide the spindle speed by the rated spindle speed to obtain the speed normalization factor; divide the feed rate by the rated feed rate to obtain the feed normalization factor; divide the depth of cut by the rated depth of cut to obtain the depth of cut normalization factor. The speed normalization factor reflects the ratio of the current spindle speed to the rated speed, the feed normalization factor reflects the ratio of the current feed rate to the rated feed rate, and the depth of cut normalization factor reflects the ratio of the current depth of cut to the rated depth of cut. All three are dimensionless. If the current spindle speed is 8000 rpm and the rated spindle speed is 10000 rpm, then the speed normalization factor is 0.8; if the current feed rate is 0.15 mm / rpm and the rated feed rate is 0.2 mm / rpm, then the feed normalization factor is 0.75; if the current depth of cut is 1.2 mm and the rated depth of cut is 1.5 mm, then the depth of cut normalization factor is 0.8.

[0033] The speed normalization factor, feed normalization factor, and depth of cut normalization factor are multiplied together, and the result is used as the working condition regularization weight factor. The basis for multiplying the three normalization factors is that cutting power is approximately proportional to the product of cutting speed, feed rate, and depth of cut. This product value comprehensively reflects the load level of the current working condition relative to the rated working condition. A value greater than 1 indicates overload machining, less than 1 indicates light load machining, and equal to 1 indicates the rated working condition. For example, multiplying the speed normalization factor (0.8), feed normalization factor (0.75), and depth of cut normalization factor (0.8) results in a product of 0.48. Therefore, the working condition regularization weight factor is 0.48, indicating that the current working condition load is lower than the rated working condition.

[0034] The gated cyclic filtering operation is specifically a weighted summation filtering based on recursive forgetting coefficients. The specific process of performing gated cyclic filtering on the transient feature map of multi-source sensors using the operating condition regularization weight factor is as follows: The gated cyclic filtering operation consists of two cascaded stages: "gating" and "cyclic filtering". The "gating" stage modulates the current feature column vector with load-sensitive amplitude using a load-condition regularization weighting factor, while the "cyclic filtering" stage performs time-domain recursive smoothing on the modulated feature column vector. After the two stages are executed sequentially, the output is a weighted feature stream that has been normalized by the load condition.

[0035] The current feature vector from the multi-source sensor transient feature map is multiplied element-wise by the load condition regularization weighting factor to obtain the gated feature vector at the current time. In signal processing, "gating" refers to the selective passage or suppression of data streams using a control signal. Here, the load condition regularization weighting factor is used as the gating signal: when the load condition regularization weighting factor is greater than 1, corresponding to high load conditions, the components of the feature vector are amplified, the gating is more open, and data from periods with abundant cutting information is enhanced and allowed to pass; when the load condition regularization weighting factor is less than 1, corresponding to low load conditions, the components of the feature vector are compressed, the gating is less open, and data from periods with scarce cutting information is suppressed and allowed to pass. This achieves differentiated weighting for different load periods in subsequent filtering.

[0036] The element-wise multiplication operation applies the working condition regularization weight factor as a scalar multiplier to each component of the feature vector. Its effect is to perform load-sensitive amplitude modulation on the feature column vector under the current working condition—the overall signal amplitude is enhanced under high load conditions, so that the time period containing rich cutting information has a higher weight in subsequent filtering; the overall signal amplitude is weakened under low load conditions, so that the contribution of the time period with scarce cutting information is suppressed.

[0037] Example 1: The aforementioned working condition regularization weight factor is 0.48. The original value of a certain component in the current feature column vector is 0.3. After element-wise multiplication, we get 0.144, indicating that the contribution of this component is reduced under low load conditions. Example 2: If the current spindle speed is 12000 rpm, the rated spindle speed is 10000 rpm, the current feed rate is 0.25 mm / rpm, the rated feed rate is 0.2 mm / rpm, the current depth of cut is 1.8 mm, and the rated depth of cut is 1.5 mm, then the speed normalization factor is 1.2, the feed normalization factor is 1.25, and the depth of cut normalization factor is 1.2. The product of these three is 1.8. The working condition regularization weight factor is 1.8. Multiplying the original value of the same component in the current feature column vector (0.3) by 1.8 yields 0.54, indicating that the contribution of this component is enhanced under high load conditions.

[0038] The gated current-time feature column vector and the filtered historical feature column vector from the previous time step are weighted and summed according to a preset recursive forgetting coefficient. The weight of the historical feature column vector is greater than the weight of the gated current-time feature column vector. The result of the weighted sum is used as the filtered output column vector for the current time step. The "cyclic filtering" stage receives the current-time feature column vector output from the "gating" stage as input and recursively weights and sums it with the filtered output from the previous time step to achieve smoothing in the time domain.

[0039] The recursive forgetting coefficient is set based on the gradual characteristics of the tool wear process, and its value ranges from greater than 0 to less than 1. The physical meaning of the recursive forgetting coefficient is the proportion of historical information in the weighted summation. The value of the recursive forgetting coefficient is based on the ratio of the tool wear time constant to the sensor signal sampling period: the tool wear time constant is obtained from tool life test data, usually ranging from 30 seconds to 300 seconds, and the sampling period is from 1 millisecond to 10 milliseconds. The ratio between the two is between 3000 and 300000. Therefore, the historical weight should be much greater than the current weight. This basis ensures that the filtered output can effectively retain the gradual wear trend while suppressing instantaneous fluctuations at the sampling level.

[0040] A typical recursive forgetting coefficient is 0.9, corresponding to a weight of 0.9 for the historical feature column vector. The weight of the gated current-moment feature column vector is 1 minus 0.9, which equals 0.1. Therefore, the filtered output column vector at the current moment is equal to 0.9 multiplied by the filtered output column vector at the previous moment plus 0.1 multiplied by the gated current-moment feature column vector. The weight of the historical feature column vector is greater than the weight of the gated current-moment feature column vector. This is based on the fact that tool wear is a slow, cumulative process, with a time constant typically ranging from several minutes to tens of minutes. However, the sampling interval of the sensor signal is on the order of milliseconds. Therefore, the filtered output at the previous moment contains long-term accumulated wear information and should be given a higher weight to maintain the continuity of the wear trend. At the same time, the current-moment feature column vector is given a lower weight to suppress high-frequency interference caused by sudden load changes.

[0041] If the recursive forgetting coefficient is 0.95, then the historical weight is 0.95 and the current weight is 0.05. In this case, the filter output relies more heavily on historical information, and the smoothing effect is more significant, making it suitable for machining scenarios where tool wear changes extremely slowly. If the recursive forgetting coefficient is 0.8, then the historical weight is 0.8 and the current weight is 0.2. In this case, the filter output responds to current information faster, making it suitable for machining scenarios where tool wear rates are relatively high.

[0042] The above operation can be viewed as a first-order recursive low-pass filter, whose cutoff frequency is determined by the recursive forgetting coefficient. The closer the recursive forgetting coefficient is to 1, the lower the cutoff frequency and the stronger the suppression capability for high-frequency interference. The filter output column vectors corresponding to all sampling times are arranged vertically in chronological order of acquisition time, forming a weighted feature stream that has undergone operating condition normalization. Each column of the weighted feature stream corresponds to a filter output column vector at a sampling time. The number of columns and rows is the same as the number of columns and rows in the multi-source sensor transient feature map, thus preserving the spatial structure of the original feature map and modulating and smoothing only the amplitude magnitude and temporal dynamic characteristics.

[0043] In the third step, the weighted feature stream serves as the input data for subsequent parallel projection operations. It possesses two characteristics: temporal weighted consistency and trend stability. Temporal weighted consistency means that the contributions of each component in the feature column vector under different load periods have been differentiated by the working condition regularization weighting factor, resulting in higher contributions during high-load periods and lower contributions during low-load periods. This allows subsequent parallel projection operations to automatically focus on periods rich in cutting information when fusing data from multiple time periods. Trend stability means that high-frequency interference caused by load abrupt changes has been suppressed by recursive filtering. The specific process of inputting the wear trend indicator into the pre-established and stored wear state assessment neural network in the CNC system, and calculating the statistical deviation between the wear trend indicator and the sample distribution under each known working condition category used during neural network training, is as follows: The wear condition assessment neural network is read from the CNC system, containing the sample center vectors and sample covariance matrices for each known working condition category stored during the training phase. The number of training samples for each known working condition category is greater than the feature vector dimension of that category. The sample center vector is the arithmetic mean of all training sample vectors within the same working condition category, representing the center position of this category in the feature space. The sample covariance matrix is ​​the average of the products of the deviation vector and its transpose for each training sample vector within the same working condition category, reflecting the correlation and dispersion range among the feature dimensions within this category. The deviation vector is the difference vector obtained by subtracting the sample center vector of this category from each training sample vector.

[0044] During the offline training phase, the number of training samples for each known working condition category is configured to be greater than the feature vector dimension of that category. Therefore, the sample covariance matrix is ​​a full-rank matrix, and its inverse matrix is ​​guaranteed to exist and be unique, making the quadratic form operation required for subsequent Mahalanobis distance calculation feasible. The current wear trend indicator is then subjected to vector difference operation with the sample center vector for each known working condition category to obtain the deviation vector of the wear trend indicator relative to each known working condition category. Each component of the deviation vector is the value of the dimension corresponding to the wear trend indicator minus the value of the dimension corresponding to the sample center vector, reflecting the direction and magnitude of the deviation of the current sample relative to the class center in each feature dimension. For each known working condition category, the deviation vector corresponding to that category is subjected to quadratic form operation with the inverse matrix of the sample covariance matrix corresponding to that category to obtain the Mahalanobis distance value for that category.

[0045] The specific calculation method for quadratic operations is as follows: First, the deviation vector is transposed to obtain its transpose. Then, this transpose is multiplied by the inverse of the sample covariance matrix, and the result is multiplied by the original deviation vector. The final result is a non-negative scalar value, which is the Mahalanobis distance. This Mahalanobis distance represents the statistical distance between the current wear trend indicator and the distribution of samples from a certain known working condition. The smaller the value, the closer the current sample is to this type of working condition distribution, and the higher the reliability; the larger the value, the farther the current sample deviates from this type of working condition distribution, and the lower the reliability. The Mahalanobis distance values ​​corresponding to all known working condition categories are then summed with weights. The weight of each Mahalanobis distance value is taken as the proportion of the number of training samples in that category to the total number of training samples. The result of the weighted sum is used as the statistical deviation between the wear trend indicator and the sample distribution of each known working condition category. During the weighted summation, categories with a larger proportion of training samples have a higher weight in the final statistical deviation, representing a higher probability of this type of working condition occurring in actual processing.

[0046] For example, if there are three known operating condition categories with 200, 300, and 500 training samples respectively, and a total of 1000 training samples, then the weights for the three categories are 0.2, 0.3, and 0.5 respectively. If the Mahalanobis distance values ​​of the current wear trend indicator relative to these three categories are 2.5, 1.2, and 3.8 respectively, then the weighted sum is 2.5 multiplied by 0.2 plus 1.2 multiplied by 0.3 plus 3.8 multiplied by 0.5, which gives 0.5 plus 0.36 plus 1.9 equals 2.76. This 2.76 is the statistical deviation.

[0047] The sample distribution is characterized by the sample center vector and sample covariance matrix under each working condition category. The specific process for calculating the statistical deviation between the wear trend indicator and the sample distribution under each known working condition category used in the neural network training is as follows: The wear condition assessment neural network is read from the CNC system, containing the sample center vectors and sample covariance matrices for each known working condition category stored during the training phase. The number of training samples for each known working condition category is greater than the feature vector dimension of that category. The sample center vector is the arithmetic mean of all training sample vectors within the same working condition category, representing the center position of this category in the feature space. The sample covariance matrix is ​​the average of the products of the deviation vector and its transpose for each training sample vector within the same working condition category, reflecting the correlation and dispersion range among the feature dimensions within this category. The deviation vector is the difference vector obtained by subtracting the sample center vector of this category from each training sample vector. The sample covariance matrix is ​​a symmetric positive definite matrix, with its diagonal elements reflecting the variance of each feature dimension and its off-diagonal elements reflecting the covariance between different feature dimensions.

[0048] During the offline training phase, the number of training samples for each known working condition category is configured to be greater than the feature vector dimension of that category, ensuring that the sample covariance matrix is ​​a full-rank matrix and its inverse matrix is ​​guaranteed to exist and be unique, thus enabling the quadratic form operation required for subsequent Mahalanobis distance calculation. The current wear trend indicator is then subjected to vector difference operation with the sample center vector for each known working condition category to obtain the deviation vector of the wear trend indicator relative to each known working condition category. Each component of the deviation vector is the value of the dimension corresponding to the wear trend indicator minus the value of the dimension corresponding to the sample center vector, reflecting the direction and magnitude of the deviation of the current sample relative to the center of that category in each feature dimension. For each known working condition category, the deviation vector corresponding to this category is subjected to quadratic form operation with the inverse matrix of the sample covariance matrix corresponding to this category to obtain the Mahalanobis distance value for that category.

[0049] The degree of statistical deviation comprehensively reflects the extent to which the current wear trend indicator deviates from the overall known working condition categories. The higher the degree of deviation, the greater the difference between the current processing condition and the known working conditions covered in the training phase. The lower the confidence of the neural network output, the more necessary it is to trigger bias correction.

[0050] When the statistical deviation exceeds a preset deviation threshold, the neural network is triggered to perform online bias correction on its output value. The specific process for obtaining the current tool wear state value after correction is as follows: The statistical deviation is compared with the deviation threshold pre-stored in the CNC system. When the statistical deviation is less than or equal to the deviation threshold, the output value of the wear state assessment neural network at the current moment is directly used as the current tool wear state value.

[0051] The deviation threshold is set based on the percentile distribution of the statistical deviation values ​​of all training samples under each known working condition category during the training phase. The specific setting method is as follows: After offline training is completed, the statistical deviation of each training sample in the training set relative to each known working condition category is calculated, resulting in a set of statistical deviation values ​​for all training samples. The 95th percentile of this set is taken as the deviation threshold. Ninety-five percent of the samples in the training set will not have a statistical deviation exceeding the 95th percentile, therefore no correction is triggered. Only the 5% of training set samples with a deviation exceeding this threshold are considered outliers, and these samples have a low probability of appearing in normal processing. For test samples with a statistical deviation exceeding this threshold, the distribution characteristics of their current wear trend indicators significantly deviate from the 95% of known working condition samples in the training set, indicating that the current processing condition may belong to a new condition not covered in the training phase, thus triggering bias correction. In the example values, if the 95th percentile of the statistical deviation values ​​for all training samples is 3.5, then the deviation threshold is set to 3.5.

[0052] When the statistical deviation exceeds the deviation threshold, the old value of the current bias term in the output layer of the wear state assessment neural network is obtained, along with the weight matrix of the output layer. Here, "the old value of the current bias term" refers to the bias parameter value stored in the output layer before the correction is triggered; subsequent update operations will use this old value as the starting point to calculate the new value. The current wear trend indicator is used as the input to the neural network. After transformation by the hidden layer of the neural network, the input feature vector of the output layer is obtained. This input feature vector is multiplied by the weight matrix and then added to the old value of the current bias term to obtain the current predicted wear value.

[0053] The arithmetic mean of the components of the wear trend indicator is taken as the expected wear value, and the difference between the predicted wear value and the expected wear value is calculated. The wear trend indicator is in vector form, and each component represents the projection value of the wear-related features extracted by the orthogonal projection operator on different channels at the same sampling time. The magnitude of each component is positively correlated with the degree of tool wear. Taking the arithmetic mean of each component yields a scalar value that comprehensively reflects the current degree of tool wear. This scalar value is used as the expected wear value to drive the direction and magnitude of the bias correction. The arithmetic mean is equal to the sum of the values ​​of each component of the wear trend indicator divided by the total number of components. This average value represents a scalar estimate of the overall strength of the wear trend indicator and is used as the expected wear value.

[0054] A step size factor is preset, with a value between 0.01 and 0.1. The step size factor is selected based on the following criteria: A value less than 0.01 results in a slow convergence speed for bias correction, requiring multiple sampling periods to approach the desired value; a value greater than 0.1 results in excessively large update steps for the bias term, potentially causing oscillations during the correction process. The range of 0.01 to 0.1 ensures that the bias correction converges to near the desired wear value at a stable and moderate speed.

[0055] The difference is multiplied by the step size factor and then added to the old value of the current bias term to obtain the updated new bias value. The accumulation operation means adding the old value of the current bias term to the product of the difference multiplied by the step size factor, and the result is used as the updated new bias value. This accumulation operation belongs to the single-step update mechanism in gradient descent, where the step size factor controls the step size of each update, and the difference serves as the approximate gradient direction of the prediction error.

[0056] The updated bias value replaces the old value of the current bias term stored in the output layer of the wear state assessment neural network, and the output value of the output layer is recalculated using the replaced network. During recalculation, the input feature vector of the hidden layer output remains unchanged; only the updated bias value is substituted into the forward calculation formula of the output layer to re-perform the weighted summation and bias addition operation. There is no need to re-execute the forward propagation of the entire neural network, thus the computational load is extremely small and can be completed within a single sampling period. This online bias correction method adjusts the output layer bias term through single-step gradient descent, enabling the network to quickly adapt to output drift under unknown working conditions while maintaining the original connection weights. The corrected output value is then used as the current tool wear state value.

[0057] The above algorithms or formulas are all dimensionless and numerical calculations, and the results are obtained by software simulation based on a large amount of collected data to obtain the most recent real-world results. The preset parameters are set by those skilled in the art according to the actual situation.

[0058] It should be understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0059] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0060] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0061] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting tool wear condition in CNC machine tools, characterized in that, Includes the following steps: The first step is to synchronously sample the raw sensing signals output by the force sensor, vibration sensor and acoustic emission sensor installed on the spindle of the CNC machine tool, and to perform short-time Fourier transform and wavelet packet decomposition on the signals of each sensor to extract the transient response components of each sensor signal in the time domain, frequency domain and time-frequency domain. All transient response components are fused and encoded into a multi-source sensing transient feature map in the order of channels. The second step is to obtain the cutting parameters of the CNC machine tool currently in operation. These cutting parameters include at least the spindle speed, feed rate, and depth of cut. Based on the cutting parameters, the working condition regularization weight factor is calculated. The working condition regularization weight factor is used to perform a gated cyclic filtering operation on the multi-source sensor transient feature map. This operation selectively enhances the slowly changing trend components in the multi-source sensor transient feature map that are positively correlated with the cumulative tool wear, while suppressing the high-frequency interference components caused by sudden load changes during the cutting process, thereby obtaining the weighted feature flow after working condition normalization. The third step is to map the weighted feature flow into two mutually orthogonal one-dimensional feature subspaces spanned by the first and second projection directions through parallel projection. The first subspace corresponding to the first projection direction is constructed to be sensitive to tool wear, and the second subspace corresponding to the second projection direction is constructed to be sensitive to instantaneous cutting load. The instantaneous cutting load component is decomposed from the weighted feature flow through this parallel projection and then removed, while retaining and outputting a pure wear trend indicator. The fourth step involves inputting the wear trend indicator into the wear state assessment neural network that has been pre-established and stored in the CNC system. The statistical deviation between the wear trend indicator and the sample distribution under each known working condition category used during the training of the neural network is calculated. When the statistical deviation exceeds the pre-set deviation threshold, the current prediction confidence is determined to have decreased, and the neural network is triggered to perform online bias correction on its output value. After correction, the current tool wear state value is obtained, and this value is output as the final detection result.

2. The method for detecting tool wear condition in CNC machine tools according to claim 1, characterized in that, The specific process of synchronous sampling is as follows: The three sensor signals are simultaneously converted from analog to digital at a preset sampling frequency to obtain three digitized raw timing signals. Perform a short-time Fourier transform on each of the original digitized time-series signals to extract the transient amplitude sequence in the time domain and the transient spectral energy distribution sequence in the frequency domain for each signal. Wavelet packet decomposition is performed on each digitized original time-series signal, and the reconstruction coefficients of each frequency band node after decomposition are used as the transient response components of the signal in the time-frequency domain. The time-domain amplitude sequence, frequency-domain energy distribution sequence, and time-frequency domain reconstruction coefficients of the force sensor channel are combined to form the feature vector of the channel. The corresponding three sequences of the vibration sensor channel are combined to form the feature vector of the channel. The corresponding three sequences of the acoustic emission sensor channel are combined to form the feature vector of the channel, resulting in three feature vectors.

3. The method for detecting tool wear condition in CNC machine tools according to claim 2, characterized in that, After obtaining three feature vectors, canonical correlation analysis is performed on the three feature vectors to solve a set of common projection directions. Each direction in the set of common projection directions corresponds to a correlation coefficient. When the number of common projection directions in the set is not less than two, the direction corresponding to the one with the largest correlation coefficient is selected as the first common projection direction, and the direction corresponding to the one with the second largest correlation coefficient is selected as the second common projection direction. The three feature vectors are projected onto the two-dimensional correlation subspace spanned by the first common projection direction and the second common projection direction, respectively, to obtain three projection vectors with the same dimension. The columns of the projection vectors are arranged in the order of sampling time. The three projection vectors are arranged vertically in a preset channel order to form a two-dimensional projection matrix, which is the multi-source sensing transient feature map.

4. The method for detecting tool wear condition in CNC machine tools according to claim 3, characterized in that, The specific process for calculating the operating condition regularization weight factor is as follows: Read the spindle speed, feed rate, and depth of cut values ​​in real time from the shared memory area of ​​the CNC system, and simultaneously read the rated spindle speed, rated feed rate, and rated depth of cut values ​​preset in the CNC system. After normalization, the speed normalization factor, feed normalization factor, and depth of cut normalization factor are obtained. The product of these three factors is used as the working condition regularization weight factor.

5. A method for detecting tool wear condition in CNC machine tools according to claim 4, characterized in that, The specific process of gated cyclic filtering is as follows: The current time feature vector in the multi-source sensor transient feature map is multiplied element by element with the operating condition regularization weight factor to obtain the gated current time feature vector. The current time feature column vector after gating is weighted and summed with the previous time filtered output historical feature column vector according to a preset recursive forgetting coefficient to obtain the current time filtered output column vector; The filtered output column vectors corresponding to all sampling times are arranged vertically in chronological order of sampling time to form a weighted feature stream that has been normalized by operating conditions.

6. A method for detecting tool wear condition in CNC machine tools according to claim 5, characterized in that, The specific process of mapping the weighted feature stream to two mutually orthogonal one-dimensional feature subspaces spanned by the first and second projection directions through parallel projection is as follows: During the offline calibration phase, historical weighted feature flow samples of CNC machine tools under various cutting conditions are collected, and the corresponding true values ​​of tool wear and instantaneous cutting load are simultaneously labeled for each historical sample. All historical samples labeled with the true value of tool wear are used to form the first sample set. The mean vector of the first sample set is calculated and used as the wear-sensitive reference center. All historical samples labeled with the true value of instantaneous cutting load are used to form a second sample set. The mean vector of this second sample set is calculated and used as the load-sensitive reference center. A connection vector is constructed with the wear-sensitive reference center as the starting point and the load-sensitive reference center as the ending point, and this connection vector is used as the first projection direction; When the wear-sensitive reference center is not a zero vector and the connecting vector is not linearly related to the wear-sensitive reference center, calculate the orthogonal complement vector of the connecting vector in the linear space where the wear-sensitive reference center is located, and use the orthogonal complement vector as the second projection direction; The first projection direction and the second projection direction are used as two mutually orthogonal projection bases for parallel projection.

7. A method for detecting tool wear condition in CNC machine tools according to claim 6, characterized in that, The specific process for retaining and outputting a clean wear trend indicator is as follows: The weighted feature stream at the current moment is projected onto the first projection direction and the second projection direction respectively to obtain the first projection coefficient along the first projection direction and the second projection coefficient along the second projection direction. Multiply the first projection coefficient by the first projection direction to reconstruct the wear-sensitive feature components at the current moment; Multiply the second projection coefficient by the second projection direction to reconstruct the load-sensitive feature components at the current moment; An orthogonal complementary projection operator is constructed based on the first projection direction and the second projection direction. The orthogonal complementary projection operator is a reprojection operator in the remaining space along the first projection direction after removing the load-sensitive feature components from the weighted feature flow. The wear-related feature vector output after the orthogonal complementary projection operator is the wear trend indicator at the current moment. The wear trend indicators at all sampling times are arranged in chronological order of collection time and then output.

8. A method for detecting tool wear condition in CNC machine tools according to claim 7, characterized in that, The sample distribution is characterized by the sample center vector and sample covariance matrix under each working condition category. The specific process for calculating the statistical deviation between the wear trend indicator and the sample distribution under each known working condition category used in the neural network training is as follows: Read the sample center vectors and sample covariance matrices for each known working condition category stored in the CNC system during the training phase of the wear condition assessment neural network; The current wear trend indicator is compared with the sample center vector of each known working condition category by performing vector difference calculation to obtain the deviation vector of the wear trend indicator relative to each known working condition category. For each known working condition category, a quadratic operation is performed between the deviation vector corresponding to that category and the inverse of the sample covariance matrix corresponding to that category to obtain the Mahalanobis distance value for that category. The Mahalanobis distance values ​​corresponding to all known working condition categories are then summed with weights, where the weight of each Mahalanobis distance value is the proportion of the number of training samples in that category to the total number of training samples. The result of the weighted sum is used as the statistical deviation between the wear trend indicator and the sample distribution under each known working condition category.

9. A method for detecting tool wear condition in CNC machine tools according to claim 8, characterized in that, When the statistical deviation exceeds a preset deviation threshold, the neural network is triggered to perform online bias correction on its output value. The specific process for obtaining the current tool wear state value after correction is as follows: When the statistical deviation is less than or equal to the deviation threshold, the output value of the wear state assessment neural network at the current moment will be directly used as the current tool wear state value output. When the statistical deviation is greater than the deviation threshold, obtain the current bias term of the output layer of the wear state assessment neural network and obtain the weight matrix of the output layer; The wear trend indicator at the current moment is used as the input of the neural network. After being transformed by the hidden layer of the neural network, the input feature vector of the output layer is obtained. The input feature vector is multiplied by the weight matrix and then added to the current bias term to obtain the current predicted wear value. The arithmetic mean of each component of the wear trend indicator is taken as the expected wear value, and the difference between the predicted wear value and the expected wear value is calculated. The difference is multiplied by a preset step size factor and then added to the current bias term to obtain the updated bias term; After replacing the current bias term in the output layer of the wear state evaluation neural network with the updated bias term, the output value of the output layer is recalculated using the replaced network and output as the current tool wear state value.

Images